Add base plays for all hosts, nvidia_gpu role, GPU pre-work items

- run.yml: base plays (geerlingguy.security) for jira and git; gpu01
  play with security + docker + nvidia_gpu
- roles/nvidia_gpu: driver pinned >=580 (Blackwell), CUDA repo,
  container toolkit incl. the nvidia-ctk runtime configure step
- manuals/20260714-nvidia-driver-install.md: dated per convention,
  corrected (pinned -server driver instead of autoinstall+cuda-drivers
  mix, toolkit optional, added missing nvidia-ctk/docker restart step)
- gpu01 folder: planning docs under notes/, runbooks under manuals/,
  scripts/; convention documented in CLAUDE.md
- scripts/share-analysis.ps1: read-only SMB share analysis for the
  Windows server (projektplan §2.1)
- TODO.md: Phase-0 pre-work items from the projektplan

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
2026-07-14 10:30:43 +02:00
co-authored by Claude Fable 5
parent 277f378fb8
commit 43054f32b8
13 changed files with 463 additions and 68 deletions
+13
View File
@@ -0,0 +1,13 @@
# TODO
## phy-z-srv-gpu01
Pre-work per [projektplan](server/phy-z-srv-gpu01/notes/20260710-projektplan.md) (§ references below):
- [x] Write a script to analyse phytrons whole file share which should be feeded to the llm: file number, structure, pdf kind, etc. → `server/phy-z-srv-gpu01/scripts/share-analysis.ps1`
- [ ] Run the share analysis on the Windows server; evaluate: corpus size (§1 revisit trigger), scanned-PDF/OCR share, index size estimate
- [ ] Storage decision (§2.3): index estimate vs. ~960 GB usable NVMe — order extra disks **before** delivery if tight
- [ ] Customer checklist (§2.2): AD bind account, read-only SMB service account, share include/exclude list, DNS name + TLS, internet access at install, chat-logging/GDPR (works council), maintenance ownership, concretize "other workloads"
- [ ] Ansible prep (§2.4): `cifs_mounts` and `llm_stack` role skeletons (`nvidia_gpu` exists); fill `group_vars/phy_z_srv_gpu01.yml` overrides; test non-GPU parts in a throwaway VM
- [ ] German eval set (§2.5): collect 2030 Q&A with the customer — becomes the acceptance criterion
- [ ] At install time: re-check and pin versions (driver / vLLM / Open WebUI / oikb / model shortlist) per re-entry checklist §5